Salesforce Inc. today announced the launch of Agentforce 3, a major upgrade to its flagship artificial intelligence product for enterprises with new ways to observe and control AI agents on the platform.

The Agentforce platform provides companies the ability to build, customize and deploy generative AI agents, which augment the work of employees autonomously. They are goal-oriented pieces of software capable of completing tasks with little or no human supervision. Using the platform, employees across sales, service, marketing and commerce can customize AI “workers” to take action on their behalf using business logic and prebuilt automations.

Today, Salesforce announced the launch of a new Command Center that provides complete observability and built-in support for the Model Context Protocol for plug-and-play compatibility with other agents and services. The company also said it is adding more than 100 new prebuilt industry actions to accelerate the deployment of industry-standard AI agents.

Command center for AI agents

As agents perform tasks behind the scenes and collaborate with human workers, there’s a growing need for improved visibility. Salesforce offers an observability platform to address technical issues that can affect the safety and performance of models. However, the new Agentforce Command Center goes a step further by unifying agent health, performance and outcome optimization.

Built into Agentforce Studio, which serves as the AI agent customization platform, the command center allows teams to analyze every agent interaction, drill into specific moments and understand trends. It will also display AI-powered recommendations for tagged conversation types to improve Agentforce agents continuously.

The command center will act as a single place to understand AI agents changing contextually according to the type of agent that is under display. For example, the metrics and graphs in the dashboard for a product delivery agent will display cancellations, shipments and other metrics. Whereas an online advertising agent will have observability for clickthrough and campaign success metrics.

Users will be able to use natural language to generate topics, instructions and test cases right in Studio. Testing Center simulates AI agent behavior at scale with data state injection and AI-driven evaluation, allowing users to stress-test agents before going live.

Tool use and agent-to-agent connectivity with MCP

Agentforce 3 now allows AI agents to connect natively to other services, tools and agents by using an open standard called Model Context Protocol, an open standard pioneered by Anthropic PBC. It allows AI agents to connect to services and other AI agents plug-and-play like a “USB-C for AI,” without the need for custom code but still governed by existing security policies.

Salesforce’s integration platform, MuleSoft, can convert any existing application programming interface into an MCP connector, complete with security policies, activity tracing and traffic controls. This will allow teams to orchestrate and govern multi-agent protocols.

For example, a user could quickly build an AI agent to use PayPal using natural language instructions to send and receive PayPal invoices using native built-in MCP integration. Using the protocol, no custom code is needed, only building the agent in the Agentforce builder interface, testing it and deploying it.

Customers will be able to connect to numerous third-party tools and resources, including Amazon Web Services, PayPal, Box, Cisco Systems, Google Cloud, IBM, Notion Labs, Stripe, Teradata and Writer.

Image: SiliconANGLE/Microsoft Designer